Self-hosted AI agent

A self-hosted AI agent is an AI agent whose app, data, and tools run on a machine you control, such as a laptop, a home server, or a VPS, instead of on a vendor's cloud. It still needs a model: a hosted one through your own API key, or a local one on your hardware.

Rakazo is an open source AI agent for persistent teammates. Each bot can use a browser and a shell, keep routines as Markdown, and pause for approval when work crosses a boundary you set. Self-hosting means that stack runs on a machine you control.

The software is Apache-2.0. Hosted Rakazo Cloud is in beta. This page is the setup you need to run it. The full self-hosting guide on GitHub covers backups, upgrades, secrets, restricted networks, and production Compose.

Diagram of a Rakazo deployment. Web, desktop, and mobile clients talk to an API. A worker and Postgres sit beside it. The computer defaults to Docker, with E2B, Daytona, CreateOS, or Boat optional. Model prompts go to the provider you configure.
A Rakazo deployment. Postgres, bot files, and the audit log stay on the machine you run. Prompts still go to the model provider you choose.

What self-hosting gives you

Data ownership

Postgres, bot files, browser profiles, and the audit log stay on the deployment you run. Content you send to a model provider is processed by that provider under its terms. You choose the provider and the keys.

Model choice

Connect OpenAI, Anthropic, Google, OpenRouter, or Vercel AI Gateway. You can also use an OpenAI-compatible endpoint or a local model server such as Ollama, LM Studio, llama.cpp, or MLX. Each bot can use a different model.

Cost control

There is no seat fee for the open source software. You pay the model provider you connect and any remote computer provider you choose. Local Docker computers are the default, so a single machine does not need an E2B, Daytona, CreateOS, or Boat account.

Local or hosted model

Self-hosting the agent does not self-host the model. With a hosted provider, the prompts and context a bot sends go to that provider. To keep them on your hardware, point the bot at a local model server such as Ollama, LM Studio, llama.cpp, or MLX, use a local approval model too, and keep the default local Docker computer.

A model server answers prompts. It is not an agent on its own. The agent is the part that keeps conversations, memory, and schedules, uses a browser and a shell, and asks before crossing a boundary you set.

Self-host only, for now

These work on your own deployment and are not yet on Rakazo Cloud. Each turns on from .env.

  • Telegram, WhatsApp, iMessage and SMS, Feishu and Lark. Add the platform's credentials: TELEGRAM_BOT_TOKEN, WHATSAPP_*, SENDBLUE_* (iMessage and SMS), or LARK_*.
  • Image and video generation. MEDIA_GENERATION_PROVIDER=fal, MEDIA_GENERATION_API_KEY, and an image or video model (MEDIA_GENERATION_IMAGE_MODEL, MEDIA_GENERATION_VIDEO_MODEL).
  • Purchases with Stripe Link. PAYMENT_WALLET_PROVIDER=link. Each purchase waits for your approval in Link.
  • Cursor cloud agents. CLOUD_AGENT_PROVIDER=cursor, CURSOR_API_KEY, and CLOUD_AGENT_SPACE_ID.
  • Local models and local MCP servers. RAKAZO_LOCAL_MODELS, and MCP_STDIO_ENABLED with MCP_STDIO_ALLOWED_COMMANDS.

The full guide lists every setting for each one.

Other self-hosted AI agents

Licenses from each project's repository on November 16, 2026. Not a ranking.
ProjectLicenseWhat you runBuilt for
RakazoApache-2.0Docker images, then the web, desktop, and mobile appsPersistent bots with routines, approvals, and an audit log
OpenClawMITA Gateway on your computer or serverAn assistant in the chat apps you already use
Hermes AgentMITA CLI, plus a gateway process for chat appsA personal agent in the terminal and in chat apps
OpenHands Agent CanvasMITA self-hosted control centerCoding agents and automations
GooseApache-2.0A desktop app, a CLI, or an APIA general agent on your own machine
n8nSustainable Use License (fair-code)Self-hosted, or n8n's cloudVisual workflows that include AI agents
LangGraphMITA library inside a program you writeStateful agents you build
CrewAIMITA Python framework inside a program you writeMulti-agent workflows you build

More detail is in Best open source AI agents in 2026. The OpenClaw comparison goes deeper on the closest one.

Requirements

  • Published images. Docker Engine 26 or newer (API 1.45 or newer, for bot home volume subpaths), the Compose plugin, curl, and OpenSSL. No Node.js install.
  • Source checkout. Node.js 24.15.0 or newer on the 24.x line, or Node.js 26 or newer, plus pnpm 12 and Docker. Node.js 25.x is not supported.
  • Hardware. A machine that runs Docker, on amd64 or arm64. Each bot computer is capped at 2 GB of memory and 2 CPU cores by default (RAKAZO_COMPUTER_MEMORY, RAKAZO_COMPUTER_CPUS). The production Compose file sizes Postgres for an 8 GB host. You need a GPU only for a local model or for GPU work inside a bot computer.

Install with published images

From an empty directory on the machine that should run Rakazo:

mkdir -p rakazo && cd rakazo &&
curl -fsSLO https://raw.githubusercontent.com/elie222/rakazo/main/infra/compose/install-images.sh &&
bash install-images.sh

The installer downloads the Compose file and .env.images.example, creates .env with random secrets, and starts Rakazo. It keeps an existing .env when you run it again. The default image tag is edge (builds from the main branch, linux/amd64 and linux/arm64).

Open http://127.0.0.1:5173 and create an account. The first registered user becomes the deployment owner. Connect a model in the app, or set OPENROUTER_API_KEY before startup. Local Docker computers are on by default.

The new bot form in Rakazo, filled with a demo bot named Inbox.
Creating a bot in the Rakazo app. The name, title, and description here are demo data.
Bot settings with the model menu open on a demo Inbox bot.
Each bot can use a different model. This menu lists models for the credentials connected on that deployment.
A routine editor for a weekday inbox check, with the instruction written as text.
A routine is readable text with a schedule. You can edit it in the app.
An approval card asking whether a demo bot may send an email, with allow once, always allow, and deny.
When work crosses a boundary you set, the bot stops on an approval card. The action is also written to the audit log.

For an agent-assisted install, use the setup prompt.

Run it on a server

Use the same installer on a VPS when bots should stay running. Images Compose binds the web app to loopback 127.0.0.1:5173. Terminate TLS on the host and proxy there. Do not publish port 3100. The web server proxies /api.

bash install-images.sh --prepare-only
# Edit .env, then:
bash install-images.sh

Before the second command, set RAKAZO_HOST to the hostname and set BETTER_AUTH_URL, WEB_ORIGIN, and API_URL to that same https:// origin. For a remote computer instead of local Docker, set SANDBOX_PROVIDER to e2b, daytona, createos, or boat and add that provider's API key.

app.example.com {
	reverse_proxy 127.0.0.1:5173
}

Replace app.example.com with your host. Desktop clients should use Existing instance with the https:// address. The public single-VM section covers production Compose, host hardening, signup allowlists, and SMTP.

If downloads are blocked, use the restricted-network guide. Secret names and recovery are in the secrets checklist.

Install from source

Use this path to develop Rakazo. The published images above are enough to run it.

git clone https://github.com/elie222/rakazo.git
cd rakazo
cp .env.example .env

Set POSTGRES_PASSWORD and POSTGRES_APP_PASSWORD to two URI-safe random values (for example openssl rand -hex 16) and put them in DATABASE_MIGRATION_URL and DATABASE_URL. Set BETTER_AUTH_SECRET, ENCRYPTION_KEY, and SCREEN_PROXY_SECRET to independent long random values. Docker sandboxes also need a dedicated SANDBOX_SUPERVISOR_TOKEN. Keep these in .env, not in git.

For host-side development with Docker Desktop, set SANDBOX_CONTROL_VIA_LOOPBACK=true in .env. The supervisor then publishes its control service on a loopback port, because Docker Desktop container addresses are not reachable from the host. Leave this unset when the supervisor runs inside Compose.

docker compose --env-file .env \
  -f infra/compose/docker-compose.yml \
  -f infra/compose/docker-compose.postgres-host.yml \
  up postgres -d
pnpm install
pnpm db:generate
pnpm db:migrate
pnpm sandbox:build
pnpm dev

The Postgres overlay publishes loopback 127.0.0.1:5433 for host-side tools. Open http://127.0.0.1:5173. docker compose down -v deletes Postgres data.

Desktop and mobile

With a server already running, the Electron app's Existing instance option takes its https:// address. HTTP is accepted only for loopback and private LAN addresses. This computer installs the published images with Docker Compose on that machine.

Run on your Mac with the desktop app. This computer installs Rakazo on that Mac. A Mac Mini can stay on as the always-on box.

In the mobile app, tap Use a custom server on the sign-in screen and enter the same HTTPS origin as WEB_ORIGIN.

Full reference

FAQ

What do I need to self-host Rakazo?

Published images need Docker Engine 26 or newer, the Compose plugin, curl, and OpenSSL. You do not need Node.js for that path. A source checkout needs Node.js 24.15.0 or newer on the 24.x line, or Node.js 26 or newer, plus pnpm 12 and Docker. Node.js 25.x is not supported.

Where does my data live?

Postgres, bot files, browser profiles, and the audit log stay on the deployment you run. Content you send to a model provider is processed by that provider under its terms. You choose the provider and the keys.

How do model choice and cost work?

You connect your own model keys, or a local or OpenAI-compatible model server, and you can pick a different model per bot. The open source software has no seat fee. You pay the model and computer providers you use. Local Docker computers are the default and do not need a hosted sandbox account.

Is hosted Rakazo available?

Yes. Rakazo Cloud is in beta at app.rakazo.com. Self-hosting is available too.

How does this compare with OpenClaw?

Both are open source and run on hardware you control. After a Docker or desktop install, Rakazo is a chat for persistent teammates. OpenClaw's getting-started guide adds a wizard, a Gateway you run in the terminal or install as a background service, and a separate step to connect a chat app.

Can a self-hosted AI agent run entirely on my own hardware?

Yes, when every model it uses is local. Rakazo with local models, including the approval model, and the default local Docker computer keeps prompts, data, and tool work on your machines. Tools that browse the web or call an integration still reach those sites.

Should I self-host or use a hosted AI agent?

Self-host when you want the data on your machines, your own model, and features that are self-host only. A hosted agent saves you upgrades, backups, and TLS. Rakazo Cloud is the hosted version of the same product, in beta.

How do I build my own AI agent?

Write one with a library such as LangGraph or CrewAI and host the program yourself, or run an agent app. In Rakazo, a new bot interviews you about the job, then keeps its routines as Markdown you can edit.

Related

Get started

Self-host with the steps above, or start on Rakazo Cloud.